EcoThink: A Green Adaptive Inference Framework for Sustainable and Accurate LLM Reasoning

Summary

EcoThink is an energy-aware adaptive inference framework for large language models (LLMs), introduced by Linxiao Li and Zhixiang Lu in an arXiv paper published on 2026-03-26. The work addresses the growing environmental footprint of LLMs as the web shifts from static retrieval to generative interaction. Current practice applies compute-intensive strategies such as chain-of-thought (CoT) reasoning indiscriminately to billions of everyday queries, causing LLMs to 'overthink'. EcoThink resolves this with a lightweight distillation-based router that dynamically assesses query complexity and routes reasoning effort accordingly. Across evaluations on nine diverse benchmarks, the framework reduces inference energy consumption by an average of 40.4%—up to 81.9% for web knowledge retrieval tasks—while incurring no statistically significant performance loss. The paper positions EcoThink as a way to reconcile high-performance AI intelligence with environmental responsibility. Paper: arXiv 2603.25498.